I made a small public repo for the boring failure modes around coding-agent work:
Repo:
It is just templates, checklists, fake scenarios, and receipts. Not claiming it solves security. It is more like a set of builder notes for slowing down at the right moments.
If you have a failure mode you think belongs in the eval scenarios, I would actually like to add it.
Hey, really cool that you put together ai-workflow-safety-kit, the four failure modes you listed (wrong folder, private context leaking, unstoppable production changes, no proof of completion) match almost exactly what I have been seeing everywhere. I am working on something in the same space and would love to compare notes sometime. What was the actual moment that pushed you to build this instead of just working around it manually?
Honestly it was a pileup of boring mistakes, not one dramatic thing: wrong repo/folder, private context drifting into public work, production/account changes getting treated like normal edits, and “done” meaning “the agent sounded confident” instead of “there is proof.”
I got tired of working around that manually, so I turned the patterns into fake scenarios and small checks people can run without handing over private logs. The Codex instruction-chain inspector came from the same itch: I wanted to know what AGENTS.md files actually load before trusting an agent run.
Would be glad to compare notes. What failure mode are you seeing most often?